"I'd like to see how they pull that off when a node goes down. I guess in "well-administered" data centers, nodes don't go down."
They offer f-fault tolerance. They can have f nodes go down in a single "zone" and keep chugging as long as no more nodes in the same zone go down before the master reconfigures. Note that the f faults are per-zone, not per-system, so in fact many more than f nodes can be down in a single system without a problem.
But, more importantly, you seem to be confusing partition tolerance and fault tolerance. CAP is about partition tolerance: offering "CA" in the presence of arbitrary partitions. They offer a specific form of fault tolerance: "CA" in the presence of any failure or partition that affects less than f nodes.
They offer f-fault tolerance. They can have f nodes go down in a single "zone" and keep chugging as long as no more nodes in the same zone go down before the master reconfigures. Note that the f faults are per-zone, not per-system, so in fact many more than f nodes can be down in a single system without a problem.
But, more importantly, you seem to be confusing partition tolerance and fault tolerance. CAP is about partition tolerance: offering "CA" in the presence of arbitrary partitions. They offer a specific form of fault tolerance: "CA" in the presence of any failure or partition that affects less than f nodes.